Right-Sizing ECG-AI for QT Safety, Ejection Fraction, and Atrial Fibrillation

Predictive Model DevelopmentPredictive External Validation

Authors:Alexis Nolin-Lapalme; Achille Sowa; Jacques Delfrate, et al.

Journal:European Society of Cardiology

Open-source ECG-AI foundation models validated across 11 cohorts deliver near-identical interpretation accuracy (micro-AUROC ~0.98–0.99), with self-supervision giving small gains for label-scarce biomarkers (LVEF≤40% AUROC 0.926 vs 0.917). For pharmacy teams, this enables scalable QT safety, GDMT, and AF-risk triage—while favoring lightweight models for routine CDS and reserving heavy SSL where benefits outweigh compute/privacy costs.

January 22, 2026

Quick Take

Two open-source foundation models—DeepECG-SL (supervised) and DeepECG-SSL (self-supervised)—show near-identical multilabel ECG interpretation performance (micro-AUROC range ≈ 0.980–0.992 across internal and external cohorts).

DeepECG-SSL shows small but statistically significant gains on selected digital biomarkers: LVEF ≤40% AUROC 0.926 vs 0.917 (Δ = 0.009, P < 0.001), 5-year incident AF AUROC 0.742 vs 0.734 (Δ = 0.008, P < 0.001), and a larger gain for LQTS genotype classification (AUROC 0.931 vs 0.850; Δ = 0.081, P = 0.026). The LQTS genotype task used relatively few labelled examples, so estimates are less precise and should be interpreted cautiously.

Operational guidance: the lightweight SL model is likely sufficient for routine ECG interpretation and low-overhead CDS; consider selectively using SSL for targeted, label-scarce screening tasks where its discrimination advantage justifies substantially higher compute, energy, and privacy costs.

Why it Matters

Inpatient ECGs are ubiquitous; ECG-derived signals such as low LVEF, LQTS genotype risk, and future AF risk map directly to medication-safety and optimisation decisions pharmacists make (e.g., QT safety reviews, eligibility for GDMT, anticoagulation triage).

Many ECG-AI tools fail to generalize across sites and devices. This study's extensive external validation across multiple vendors and healthcare systems is an important step toward generalizability, but local validation and calibration remain required before clinical use.

Deployment must balance clinical value against compute, energy, privacy (membership-inference) and alert-fatigue risks. Safe adoption will need governance, local threshold tuning, worklist-driven stewardship, and human-in-the-loop review rather than autonomous actioning.

What They Did

Retrospective multi-center development using the Montreal Heart Institute archive (~1.45M ECGs) with patient-level 70/10/20 splits; DeepECG-SSL pre-training pooled unlabelled ECGs from MHI, Code-15 (≈345k ECGs) and MIMIC-IV (≈558k ECGs) before fine-tuning on labelled tasks.

Built and released two open-source foundation models: DeepECG-SL (supervised multilabel model trained on 77 diagnostic labels) and DeepECG-SSL (Wav2Vec + transformer self-supervised pre-training then fine-tuned for ECG interpretation and digital biomarkers); applied a standardized three-step ECG preprocessing pipeline to harmonize signals across vendors and published Dockerized code and weights.

Performed extensive external validation across 11 cohorts (~881k ECGs), plus age/sex fairness audits, membership-inference privacy testing, LIME/Grad-CAM explainability checks, decision-threshold tuning on validation splits, and documented computational and energy requirements (noting SSL's substantially larger footprint).

What They Found

Both models achieved near-identical multilabel ECG-interpretation performance (internal micro-AUROC 0.990–0.992; external micro-AUROC 0.980–0.983), with only minimal differences across common diagnostic categories.

DeepECG-SSL outperformed DeepECG-SL on selected digital biomarkers: LVEF ≤40% AUROC 0.926 vs 0.917 (Δ = 0.009, P < 0.001; tested on ~25k paired ECG–echocardiogram samples), 5-year incident AF AUROC 0.742 vs 0.734 (Δ = 0.008, P < 0.001; tested on ~132k ECGs), and a larger gain for LQTS genotype classification (AUROC 0.931 vs 0.850; Δ = 0.081, P = 0.026). The genotype/subtyping result was derived from a small labelled set and therefore has wider uncertainty.

The SSL performance advantage widened as labelled fine-tuning data decreased, with the largest relative gains on rare/low-n tasks (e.g., LQTS subtyping), demonstrating label efficiency for targeted screening.

Trade-offs are substantial: DeepECG-SSL is much larger and costlier (≈90.37M vs ≈1.51M parameters; ~27–60× more compute depending on metric; energy per 1,000 ECGs 0.7463 vs 0.1786 Wh). Fairness audits reported small TPR/FPR disparities across age and sex (ΔTPR < 0.1, ΔFPR < 0.02), but both models showed susceptibility to membership-inference attacks on some external cohorts (re-identification AUROC > 0.95), raising privacy concerns for deployment without safeguards.

Takeaways

Treat ECG-AI outputs as prioritization/screening tools, not autonomous decision-makers: use model signals to focus QT-safety checks, identify likely low-LVEF candidates for GDMT review, and triage AF-risk follow-up while requiring clinician confirmation.

Interpret the SL vs SSL trade-off as scope-setting: SL is cost-efficient for routine interpretations and low-overhead CDS; SSL adds value for label-scarce biomarkers and rare-subtype classification but should be reserved for targeted use cases where the incremental clinical benefit outweighs compute, privacy, and monitoring costs.

Before deployment mandate local validation and threshold tuning, explicit governance for privacy and alerting, human-in-the-loop workflows to reduce alert fatigue, and operational plans for monitoring calibration, drift, fairness, and membership-inference risk.

Strengths and Limitations

Strengths:

Extensive external validation across 11 datasets (~881k ECGs) and a standardized preprocessing pipeline strengthen cross-site generalizability claims.

Open-source release of model weights, preprocessing code, and Dockerized deployment plus published fairness and privacy analyses support reproducibility and enable local validation and integration.

Limitations:

Labelled training originates primarily from a single tertiary centre (MHI) and a retrospective, adult-predominant cohort; pediatric and broader sociodemographic validation is limited.

Several rare-task results (notably LQTS genotype/subtyping) were derived from relatively small labelled sets, increasing uncertainty. The source manuscript contains inconsistent reporting of sample counts for some LQTS analyses; confirm the original dataset counts before operational use.

Both models (particularly SSL) exhibited elevated membership-inference vulnerability on some external cohorts and SSL's substantially larger compute and energy footprint constrains routine real-time deployment without privacy safeguards and ML-ops resources.

Bottom Line

Self-supervised ECG foundation models show promise for screening and targeted, label-scarce biomarker tasks while supervised models are likely sufficient and more efficient for routine ECG interpretation; prioritize local validation, thresholding, governance for privacy and alerting, and selective use of SSL where clinical benefit justifies the resource and privacy costs.